feat(musa): 13 reduction/normalization mudnn kernels + unbreak MUSA build

[lvyufeng]

Summary

Second mudnn batch: 13 reduction/normalization operators, taking MUSA coverage
92 -> 105 kernels. Follows #60 and uses the same table-driven route -- 8 new
templates in scripts/codegen_mudnn.py, no framework changes.

This re-targets main. #61 was opened against #60's branch, and #60 was
squash-merged, so #61's content never reached main -- these commits are rebased onto
main instead.

New ops:

• softmax family: _log_softmax _softmax_backward_data _log_softmax_backward_data
• reductions: amax amin prod.dim_int any.dim all.dim
• statistics: var.correction std.correction linalg_vector_norm
• normalization: native_layer_norm native_layer_norm_backward

Softmax::RunBwd's operands are (gradInput, output, gradOutput), measured against
CPU, and the numbers match aten's y*(g - sum(g*y)) exactly. input_dtype only
records what the forward input was; when it differs from the gradient's dtype aten
wants a converting backward, which mudnn does not express, so that combination goes
to the host rather than silently computing in the wrong precision.

The Reduce fix (2nd commit)

The existing guard only covered multi-dim reductions, where mudnn v3300 raises
SIGFPE on an input that is a broadcast of one element. A single-dim reduce over
the same input has a second, quieter failure: it intermittently writes only out[0]
and leaves the rest of the output untouched, so the answer is whatever the caching
allocator last left in that block.

It surfaces as a wrong bias gradient. linear(x, w, b).sum().backward() reduces a
grad_output that autograd builds as ones.expand(...), whose storage is a single
float, and the gradient came back as [4, 34, 38, 42, 46] -- element 0 correct,
elements 1.. left over from a previous op.

Two properties worth knowing before reviewing:

• It does not reproduce outside the allocator. Probed standalone against a
  deliberately poisoned output buffer, mudnn is correct every time. The trigger needs
  a recycled non-zero block underneath, so a C++ probe will report the library fine.
• A pure-tensor reduce cannot serve as the regression test, for the same reason.
  The test added here goes through the bias-gradient path, which is what actually
  leaves recycled data in the output block.

MudnnReduceWouldFault becomes MudnnReduceNeedsContiguous, dropping the dim-count
argument and keying only on "fully 0-strided". The materialization it forces reads one
element, so the cost is negligible.

The guard.h fix (3rd commit)

Unrelated to the kernels, but MUSA cannot build without it. #55 added a gate around
#include <c10/cuda/CUDAStream.h> in csrc/runtime/guard.h listing ASCEND,
TSINGMICRO, DCU and GCU -- but not MUSA. The Moore Threads toolkit ships no CUDA
runtime, so the header is absent and every MUSA build has failed since that merge
with cuda_runtime_api.h: No such file or directory while compiling guard.cc.
Confirmed by building flagos/main unmodified on an MTT S5000: it fails, and builds
with this one-line change. hooks.h and copy_ops.cc already carry
!defined(USE_MUSA) on the equivalent gate.

Test plan

Environment: torch==2.10.0+cpu, 8x MTT S5000, --no-build-isolation.

• Numerics vs CPU for all 13 ops across float32/float16, including keepdim both
  ways, negative dims, multi-dim reductions and correction=0/1 for var/std.
• Reduce regression: the bias-gradient path, repeated after ops that dirty the
  allocator, now returns [4, 4, 4, 4, 4] every time. Added as a test.
• Suite: pytest tests/ --ignore=tests/manual/metax gives 814 passed / 203
  skipped / 4 xpassed / 24 failed. The same run on flagos/main at a02f62b (plus
  only the guard.h fix, since main otherwise does not build here) gives 810 passed /
  24 failed -- the failure sets are byte-identical, so passed goes up by 4 and
  nothing regresses. Those 24 are pre-existing environment limits on this box:
  profiler tests need CUPTI, test_compile.py needs inductor's CUDA backend. The 11
  errors are the known missing transformers.
• tests/manual/metax is excluded because test_qwen3_fsdp2_metax.py (added by feat(metax): FSDP2 full feature parity + Qwen3 training match #62)
  runs argparse at import time and consumes pytest's own argv, which aborts
  collection for the whole run with INTERNALERROR> SystemExit: 2. That is unrelated
  to this PR and reproduces on unmodified main.
• ruff==0.15.12 check + format clean.
• Regenerating with python scripts/codegen_mudnn.py at each commit reproduces the
  checked-in generated files exactly.

🤖 Generated with Claude Code